Artificial intelligence-derived quantitative blastocyst morphology for objective embryo assessment and fetal heart tone stratification

Morphology-based predicted grades showed higher agreement with consensus grades than individual embryologists for developmental stage and inner cell mass. Prior work on objectifying blastocyst grading often relied on end-to-end whole-image prediction; this study instead segments the zona pellucida, inner cell mass, and trophectoderm first and derives indicators from the resulting masks, linking model outputs to established embryological structures. Weighted kappa comparison across 10,718 embryos: predicted grades for developmental stage reached 0.865 (almost perfect) versus 0.605 for embryologists (substantial); for inner cell mass, 0.730 versus 0.610; for trophectoderm, the predicted mean was slightly lower (0.609 versus 0.714) but with lower variability across comparisons.

A Random Forest model built on quantitative morphological indicators outperformed a manual consensus grade-based model for fetal heart tone prediction. Most previous AI embryo assessment focused on whole-image discriminative performance; this study directly used a manual consensus grade-based model as the comparator and compared AUROCs with DeLong's method. 1,387 full, expanded, or hatching blastocysts from six centers; out-of-fold AUROC under nested cross-validation (five outer folds, three inner splits) was 0.648 versus 0.610, DeLong's test p = 0.04; both models exceeded chance (p < 0.001).

A surrogate CART model showed a wider stratification of observed fetal heart tone rates for quantitative morphology, with key node indicators consistent with SHAP results. Rather than reporting AUROC alone, the study used an interpretable surrogate tree to translate predicted probabilities into observed fetal heart tone rates at terminal nodes and compared it with the consensus grade-based model under identical conditions. The surrogate CART achieved a fidelity R² of 0.859; observed fetal heart tone rates ranged from 11.8% to 46.8% (range 35.0%) for the quantitative morphology model versus 16.7% to 36.2% (range 19.5%) for the consensus grade model; SHAP identified maximum zona thickness (mean absolute SHAP value 0.029), peripheral cell count (0.021), and ICM compactness (0.020) as most influential.

Seventeen indicators derived from segmentation masks were expressed as relative or ratio-based measures and covered embryo, inner cell mass, trophectoderm, and zona pellucida levels. This indicator set converts morphological features otherwise assessed qualitatively, such as compactness, cell number, and zona thickness, into reproducible numeric quantities and aims to reduce the influence of image scale and equipment differences. Indicators comprise one embryo-level (expansion index), six ICM-related, four TE-related, and six zona pellucida-related measures, extracted with libraries including OpenCV, scikit-image, alphashape, shapely, pyradiomics, and SimpleITK.

The results pertain to full, expanded, and hatching blastocysts relevant to day 5 transfer in routine IVF laboratories, with indicators expressed in relative or ratio form to facilitate comparison across microscope settings and equipment; for readers, this means quantitative morphology can serve as an objective complement to manual grading, useful for reviewing developmental stage and inner cell mass grades and for providing interpretable morphology-level stratification related to fetal heart tone.

The fetal heart tone analysis was restricted to full, expanded, and hatching blastocysts, so applicability to early and fully hatched blastocysts remains unclear; fetal heart tone is not equivalent to live birth, and the relationship to ongoing pregnancy and live birth remains to be evaluated; the model did not incorporate maternal age, endometrial thickness, ovarian stimulation characteristics, hormonal status, or transfer-related factors; model agreement for trophectoderm grading was slightly lower than that of individual embryologists, possibly related to the difficulty of representing trophectoderm morphology distributed along the blastocyst circumference within a single two-dimensional focal plane; and the cohort is single-country Korean data, so generalization to other populations and clinical settings requires external validation. In addition, the text available for this reading was a fast parse in which figures and supplementary tables were not fully rendered, so specific node-level values for Figure 2, Figure 3, and Supplementary Tables S1-S3 are reported here only as described in the main text.